Seminar | March 6 | 3 p.m. | 8013 Berkeley Way West
Bill Peebles, UC Berkeley
Electrical Engineering and Computer Sciences (EECS)
Large-scale generative models have fueled recent progress in artificial intelligence, especially within NLP. In this talk, I'll present ways to train improved and scalable generative models of images and other modalities beyond language. I'll begin by introducing a new, powerful class of generative models---Diffusion Transformers (DiTs). With one small yet critically-important architecture modification, Transformers retain their excellent scaling properties as diffusion models and outperform all prior convolution-based U-Net models which have previously dominated image synthesis. Next, I'll introduce a novel framework for learning to learn based on generative models of a new data source---neural network checkpoints. I'll show that diffusion models can learn to optimize by generating trained neural network parameters, conditioned on a target loss or reward. Finally, I'll discuss how pre-trained image-level generative models can be used to tackle a downstream task in vision without requiring any task-specific training data. I'll show how pre-trained GAN generators can synthesize paired data to train networks for the dense visual correspondence problem---without requiring any human-annotated supervision or even real images.
Jean Nguyen, jeannguyen@eecs.berkeley.edu, 510-643-8347